GitLab 16 changes

Tier: Free, Premium, Ultimate Offering: Self-managed

This page contains upgrade information for minor and patch versions of GitLab 16. Ensure you review these instructions for:

  • Your installation type.
  • All versions between your current version and your target version.

For more information about upgrading GitLab Helm Chart, see the release notes for 7.0.

Issues to be aware of when upgrading from 15.11

  • PostgreSQL 12 is not supported starting from GitLab 16. Upgrade PostgreSQL to at least version 13.6 before upgrading to GitLab 16.0 or later.
  • If your GitLab instance upgraded first to 15.11.0, 15.11.1, or 15.11.2 the database schema is incorrect. Perform the workaround before upgrading to 16.x.
  • Starting with 16.0, GitLab self-managed installations now have two database connections by default, instead of one. This change doubles the number of PostgreSQL connections. It makes self-managed versions of GitLab behave similarly to GitLab.com, and is a step toward enabling a separate database for CI features for self-managed versions of GitLab. Before upgrading to 16.0, determine if you need to increase max connections for PostgreSQL.

  • Most installations can skip 16.0, 16.1, and 16.2, as the first required stop on the upgrade path is 16.3. In all cases, you should review the notes for those intermediate versions.

    Some GitLab installations must stop at those intermediate versions depending on which features are used and the size of the environment:

    • 16.0.8: Instances with lots of records in the users table. For more information, see long-running user type data change.
    • 16.1.5: Instances that use the NPM package registry.
    • 16.2.8: Instances with lots of pipeline variables (including historical pipelines).

    If your instance is affected and you skip these stops:

    • The upgrade can take hours to complete.
    • The instance generates 500 errors until all the database changes are finished, after which Puma and Sidekiq must restarted.
    • For Linux package installations, a timeout occurs and a manual workaround to complete the migrations is necessary.

  • GitLab 16.0 introduced changes around enforcing limits on project sizes. On self-managed, if you use these limits, projects that have reached their limit causes error messages when pushing to unaffected Git repositories in the same group. The errors often refer to exceeding a limit of zero bytes (limit of 0 B).

    The pushes succeed, but the errors imply otherwise and might cause issues with automation. Read more in the issue. The bug is fixed in GitLab 16.5 and later.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

16.11.0

  • A groups_direct field was added to the JSON web token (ID token).
    • If you use GitLab CI/CD ID tokens to authenticate with third party services, this change can cause the HTTP header size to increase. Proxy servers might reject the request if the headers get too big.
    • If possible, increase the header limit on the receiving system.
    • See issue 467253 for more details.
  • After upgrading to GitLab 16.11 some users with large environments and databases experience timeouts loading source code pages in the web UI.
    • These timeouts are caused by slow PostgreSQL queries for pipeline data, which then exceed the internal 60 second timeout.
    • You can still clone Git repositories, and other requests for repository data works.
    • See issue 472420 for more details, including steps to confirm you’re affected and housekeeping to run in PostgreSQL to correct it.

Linux package installations

In GitLab 16.11, PostgreSQL is automatically upgraded to 14.x except for the following cases:

  • You are running the database in high availability using Patroni.
  • Your database nodes are part of a GitLab Geo configuration.
  • You have specifically opted out from automatically upgrading PostgreSQL.
  • You have postgresql['version'] = 13 in your /etc/gitlab/gitlab.rb.

Fault-tolerant and Geo installations support manual upgrades to PostgreSQL 14, see Packaged PostgreSQL deployed in an HA/Geo Cluster.

Geo installations

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

  • In GitLab 16.11 through GitLab 17.2, a missing PostgreSQL index can cause high CPU usage, slow job artifact verification progress, and slow or timed out Geo metrics status updates. The index was added in GitLab 17.3. To manually add the index, see Geo Troubleshooting - High CPU usage on primary during job artifact verification.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.11 All None
    17.0 All None
    17.1 All None
    17.2 All None

16.10.0

You might encounter the following error while upgrading to GitLab 16.10 or later: 

PG::UndefinedColumn: ERROR:  column namespace_settings.delayed_project_removal does not exist

This error can occur when a migration that removes the column runs before a later migration runs that references the now-deleted column. A fix for this bug is planned for release in 16.11.

To workaround the problem:

  1. Temporarily re-create the column. Using gitlab-psql or connecting to the database manually, run: 

    ALTER TABLE namespace_settings ADD COLUMN delayed_project_removal BOOLEAN DEFAULT NULL;

  2. Apply pending migrations: 

    gitlab-ctl reconfigure

  3. Finalize checks: 

    gitlab-ctl upgrade-check

  4. Remove the column. Using gitlab-psql or connecting to the database manually, run: 

    ALTER TABLE namespace_settings DROP COLUMN delayed_project_removal;

Linux package installations

Linux package installations for GitLab 16.10 include an upgrade to a new major version of Patroni, from version 2.1.0 to version 3.0.1.

If you’re using one of the reference architectures that enables High Availability (HA) (3k users or more), you’re using PostgreSQL replication and failover for Linux package installations, which uses Patroni.

If this is your case, read Multi-node upgrades with downtime on how to upgrade your multi-node instance.

For more information on the changes introduced between version 2.1.0 and version 3.0.1, see the Patroni release notes.

Geo installations

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

16.9.0

You might encounter the following error while upgrading to GitLab 16.9.0: 

PG::UndefinedTable: ERROR:  relation "p_ci_pipeline_variables" does not exist

Make sure that all migrations complete and restart all Rails and Sidekiq nodes. A fix for this bug is planned for release in 16.9.1.

Geo installations

  • Due to a bug in container replication, a misconfigured secondary could mark a failed container replication as successful. Subsequent verification would mark the container as failed due to a checksum mismatch. The workaround is to fix the secondary configuration. Affected releases:

    Affected minor releases Affected patch releases Fixed in
    All All 16.10.2

  • Due to a bug introduced GitLab 16.5, personal snippets are not being replicated to secondary Geo sites. This can lead to loss of personal snippet data in the event of a Geo failover. See details of the problem and workaround in issue #439933.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 All None
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • You might experience verification failures on a subset of container registry images due to checksum mismatch between the primary site and the secondary site. Issue 442667 describes the details. While there is no direct risk of data loss as the data is being correctly replicated to the secondary sites, it is not being successfully verified. There are no known workarounds at this time.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

Linux package installations

  • The Sidekiq min_concurrency and max_concurrency options are deprecated in GitLab 16.9.0 and due for removal in GitLab 17.0.0. In GitLab 16.9.0 and later, to avoid breaking changes in GitLab 17.0.0, set the new concurrency option and remove the min_concurrency and max_concurrency options.

16.8.0

  • In GitLab 16.8.0 and 16.8.1, the Sidekiq gem was upgraded, and the newer version requires Redis 6.2 or later. If you are using Redis 6.0, upgrade directly to 16.8.2, which restores compatibility with Redis 6.0.
  • note
    You should upgrade to Redis 6.2 or later as Redis 6.0 is no longer supported.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Geo installations

  • PostgreSQL version 14 is the default for fresh installations of GitLab 16.7 and later. Due to a known issue, existing Geo secondary sites cannot be upgraded to PostgreSQL version 14. For more information, see issue 7768. All Geo sites must run the same version of PostgreSQL. To add a new Geo secondary site on GitLab 16.7 to 16.8.1, you must take one of the following actions based on your configuration:

    • To add your first Geo secondary site: Upgrade the Primary site to PostgreSQL 14 before you set up the new Geo secondary site. No special action is required if your primary site is already running PostgreSQL 14.
    • To add a new Geo secondary site to a deployment that already has one or more Geo secondaries:
      • If all existing sites are running PostgreSQL 13, install the new Geo secondary site with pinned PostgreSQL version 13.
      • If all existing sites are running PostgreSQL 14: No special action is required.
      • Upgrade all existing sites to GitLab 16.8.2 or later and PostgreSQL 14 before you add the new Geo secondary site to the deployment.

  • Due to a bug introduced GitLab 16.5, personal snippets are not being replicated to secondary Geo sites. This can lead to loss of personal snippet data in the event of a Geo failover. See details of the problem and workaround in issue #439933.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 All None
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • You might experience verification failures on a subset of container registry images due to checksum mismatch between the primary site and the secondary site. Issue 442667 describes the details. While there is no direct risk of data loss as the data is being correctly replicated to the secondary sites, it is not being successfully verified. There are no known workarounds at this time.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

16.7.0

  • GitLab 16.7 is a required upgrade stop. This ensures that all database changes introduced in GitLab 16.7 and earlier have been implemented on all self-managed instances. Dependent changes can then be released in GitLab 16.8 and later. Issue 429611 provides more details.

    • If you skip 16.6 in your upgrade path, you might experience performance issues after upgrading to 16.7 when your instance processes a background database migration from the GitLab 16.6 release. Read more about the ci_builds migration in the 16.6.0 upgrade notes.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

Specific information applies to Linux package installations:

  • As of GitLab 16.7, PostgreSQL 14 is the default version installed with the Linux package. During a package upgrade, the database isn’t upgraded to PostgreSQL 14. If you want to upgrade to PostgreSQL 14, you must do it manually.

    If you want to use PostgreSQL 13, you must set postgresql['version'] = 13 in /etc/gitlab/gitlab.rb.

Geo installations

  • PostgreSQL version 14 is the default for fresh installations of GitLab 16.7 and later. Due to a known issue, existing Geo secondary sites cannot be upgraded to PostgreSQL version 14. For more information, see issue. All Geo sites must run the same version of PostgreSQL. To add a new Geo secondary site based on GitLab 16.7 to 16.8.1, you must take one of the following actions based on your configuration:

    • You are adding your first Geo secondary site: Upgrade the Primary site to PostgreSQL 14 before setting up the new Geo secondary site. No special action is required if your primary site is already running PostgreSQL 14.
    • You are adding a new Geo secondary site to a deployment that already has one or more Geo secondaries:
      • If all existing sites are running PostgreSQL 13: Install the new Geo secondary site with pinned PostgreSQL version 13.
      • If all existing sites are running PostgreSQL 14: No special action is required.
      • Upgrade all existing sites to GitLab 16.8.2 or later and PostgreSQL 14 before you add the new Geo secondary site to the deployment.

  • You might experience verification failures on a subset of projects due to checksum mismatch between the primary site and the secondary site. The details are tracked in this issue. There is no risk of data loss as the data is being correctly replicated to the secondary sites. Users cloning impacted projects from a Geo secondary site will always be redirected to the primary site. There are no known workarounds at this time. We are actively working on a fix.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 16.6.0 - 16.6.5 16.6.6
    16.7 16.7.0 - 16.7.3 16.7.4

  • Due to a bug introduced GitLab 16.5, personal snippets are not being replicated to secondary Geo sites. This can lead to loss of personal snippet data in the event of a Geo failover. See details of the problem and workaround in issue #439933.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 All None
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

16.6.0

  • GitLab 16.6 introduces a background migration that re-writes every row in the CI jobs table (ci_builds) as part of upgrading the primary key to 64 bits. ci_builds is one of the largest tables on most GitLab instances, so this migration runs more aggressively than usual to ensure it takes a reasonable amount of time. Background migrations usually pause between batches of rows, but this migration does not.

    This might cause performance issues in self-managed environments:

    • Disk I/O will be higher than usual. This will be a particular issue for instances hosted by cloud providers where disk I/O is restricted.
    • Autovacuum might run more frequently in the background to ensure the old rows (dead tuples) are removed, and to perform other related housekeeping.
    • Queries might run slowly, temporarily, because inefficient query plans get selected by PostgreSQL. This might be triggered by the volume of change on the table.

    Workarounds:

    • Pause the running migration in the Admin area.

    • Recreate table statistics manually on the database console to ensure the right query plan is selected: 

      SET statement_timeout = 0;
VACUUM FREEZE VERBOSE ANALYZE public.ci_builds;
  • Old CI Environment destroy jobs may be spawned after upgrading to GitLab 16.6.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Geo installations

  • You might experience verification failures on a subset of projects due to checksum mismatch between the primary site and the secondary site. The details are tracked in this issue. There is no risk of data loss as the data is being correctly replicated to the secondary sites. Users cloning impacted projects from a Geo secondary site will always be redirected to the primary site. There are no known workarounds at this time. We are actively working on a fix.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 16.6.0 - 16.6.5 16.6.6
    16.7 16.7.0 - 16.7.3 16.7.4

  • Due to a bug introduced GitLab 16.5, personal snippets are not being replicated to secondary Geo sites. This can lead to loss of personal snippet data in the event of a Geo failover. See details of the problem and workaround in issue #439933.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 All None
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

16.5.0

  • Git 2.42.0 and later is required by Gitaly. For self-compiled installations, you should use the Git version provided by Gitaly.
  • A regression may sometimes cause an HTTP 500 error when navigating a group. Upgrading to GitLab 16.6 or later resolves the issue.
  • A regression may cause Unselected Advanced Search facets to not load. Upgrading to 16.6 or later resolves the issue.

  • The unique_batched_background_migrations_queued_migration_version index was introduced in 16.5, and the post deployment migration DeleteOrphansScanFindingLicenseScanningApprovalRules2 has the potential to break this unique constraint while doing a zero-downtime upgrade. A workaround is available in issue #437291 which fixes the error: 

    PG::UniqueViolation: ERROR:  duplicate key value violates unique constraint
"unique_batched_background_migrations_queued_migration_version"
DETAIL:  Key (queued_migration_version)=(20230721095222) already exists.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

  • SSH clone URLs can be customized by setting gitlab_rails['gitlab_ssh_host'] in /etc/gitlab/gitlab.rb. This setting must now be a valid hostname. Previously, it could be an arbitrary string that was used to show a custom hostname and port in the repository clone URL.

    For example, prior to GitLab 16.5, the following setting worked: 

    gitlab_rails['gitlab_ssh_host'] = "gitlab.example.com:2222"

    Starting with GitLab 16.5, the hostname and port must be specified separately: 

    gitlab_rails['gitlab_ssh_host'] = "gitlab.example.com"
gitlab_rails['gitlab_shell_ssh_port'] = 2222

    After you change the setting, make sure to reconfigure GitLab: 

    sudo gitlab-ctl reconfigure

Geo installations

Specific information applies to installations using Geo:

  • A number of Prometheus metrics were incorrectly removed in 16.3.0, which can break dashboards and alerting:

    Affected metric Metric restored in 16.5.2 and later Replacement available in 16.3+
    geo_repositories_synced Yes geo_project_repositories_synced
    geo_repositories_failed Yes geo_project_repositories_failed
    geo_repositories_checksummed Yes geo_project_repositories_checksummed
    geo_repositories_checksum_failed Yes geo_project_repositories_checksum_failed
    geo_repositories_verified Yes geo_project_repositories_verified
    geo_repositories_verification_failed Yes geo_project_repositories_verification_failed
    geo_repositories_checksum_mismatch No None available
    geo_repositories_retrying_verification No None available
    • Impacted versions:
      • 16.3.0 to 16.5.1
    • Versions containing fix:
      • 16.5.2 and later

    For more information, see issue 429617.

  • Object storage verification was added in GitLab 16.4. Due to an issue some Geo installations are reporting high memory usage which can lead to the GitLab application on the primary becoming unresponsive.

    Your installation may be impacted if you have configured it to use object storage and have enabled GitLab-managed object storage replication

    Until this is fixed, the workaround is to disable object storage verification. Run the following command on one of the Rails nodes on the primary site: 

    sudo gitlab-rails runner 'Feature.disable(:geo_object_storage_verification)'

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.4 16.4.0 - 16.4.2 16.4.3
    16.5 16.5.0 - 16.5.1 16.5.2

  • After Group Wiki verification was added in GitLab 16.3, missing Group Wiki repositories are being incorrectly flagged as failing verification. This issue is not a result of an actual replication/verification failure but an invalid internal state for these missing repositories inside Geo and results in errors in the logs and the verification progress reporting a failed state for these Group Wiki repositories.

    See details of the problem and workaround in issue #426571

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 16.5.0 - 16.5.1 16.5.2

  • You might experience verification failures on a subset of projects due to checksum mismatch between the primary site and the secondary site. The details are tracked in this issue. There is no risk of data loss as the data is being correctly replicated to the secondary sites. Users cloning impacted projects from a Geo secondary site will always be redirected to the primary site. There are no known workarounds at this time. We are actively working on a fix.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 16.6.0 - 16.6.5 16.6.6
    16.7 16.7.0 - 16.7.3 16.7.4

  • Due to a bug introduced GitLab 16.5, personal snippets are not being replicated to secondary Geo sites. This can lead to loss of personal snippet data in the event of a Geo failover. See details of the problem and workaround in issue #439933.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 All None
    16.8 16.8.0 - 16.8.3 16.8.4
    16.9 16.9.0 - 16.9.1 16.9.2

  • Due to a bug introduced GitLab 16.5 and fixed in 17.0, GitLab Pages deployment files are being orphaned on secondary Geo sites. If Pages deployments are stored locally, then this can lead to zero remaining storage and subsequently data loss in the event of a failover. See details of the problem and workaround in issue #457159.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.7 16.7.8
    16.8 16.8.0 - 16.8.7 16.8.8
    16.9 16.9.0 - 16.9.8 16.9.9
    16.10 16.10.0 - 16.10.6 16.10.7
    16.11 16.11.0 - 16.11.3 16.11.4

16.4.0

  • Updating a group path received a bug fix that uses a database index introduced in 16.3.

    If you upgrade to 16.4 from a version lower than 16.3, you must execute ANALYZE packages_packages; in the database before you use it.

  • You might encounter the following error while upgrading to GitLab 16.4 or later: 

    main: == 20230830084959 ValidatePushRulesConstraints: migrating =====================
main: -- execute("SET statement_timeout TO 0")
main:    -> 0.0002s
main: -- execute("ALTER TABLE push_rules VALIDATE CONSTRAINT force_push_regex_size_constraint;")
main:    -> 0.0004s
main: -- execute("RESET statement_timeout")
main:    -> 0.0003s
main: -- execute("ALTER TABLE push_rules VALIDATE CONSTRAINT delete_branch_regex_size_constraint;")
rails aborted!
StandardError: An error has occurred, all later migrations canceled:

PG::CheckViolation: ERROR:  check constraint "delete_branch_regex_size_constraint" of relation "push_rules" is violated by some row

    These constraints might return an error:

    • author_email_regex_size_constraint
    • branch_name_regex_size_constraint
    • commit_message_negative_regex_size_constraint
    • commit_message_regex_size_constraint
    • delete_branch_regex_size_constraint
    • file_name_regex_size_constraint
    • force_push_regex_size_constraint

    To fix the error, find the records in the push_rules table that exceed the 511 character limit. 

    ;; replace `delete_branch_regex` with a name of the field used in constraint
SELECT id FROM push_rules WHERE LENGTH(delete_branch_regex) > 511;

    To find out if a push rule belongs to a project, group, or instance, run this script in the Rails console: 

    # replace `delete_branch_regex` with a name of the field used in constraint
long_rules = PushRule.where("length(delete_branch_regex) > 511")

array = long_rules.map do |lr|
  if lr.project
    "Push rule with ID #{lr.id} is configured in a project #{lr.project.full_name}"
  elsif lr.group
    "Push rule with ID #{lr.id} is configured in a group #{lr.group.full_name}"
  else
    "Push rule with ID #{lr.id} is configured on the instance level"
  end
end

puts "Total long rules: #{array.count}"
puts array.join("\n")

    Reduce the value length of the regex field for affected push rules records, then retry the migration.

    If you have too many affected push rules, and you can’t update them through the GitLab UI, contact GitLab support.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Self-compiled installations

  • A new method of configuring paths for the GitLab secret and custom hooks is preferred in GitLab 16.4 and later:
    1. Update your configuration [gitlab] secret_file to configure the path to the GitLab secret token.
    2. If you have custom hooks, update your configuration [hooks] custom_hooks_dir to configure the path to server-side custom hooks.
    3. Remove the [gitlab-shell] dir configuration.

Geo installations

Specific information applies to installations using Geo:

  • A number of Prometheus metrics were incorrectly removed in 16.3.0, which can break dashboards and alerting:

    Affected metric Metric restored in 16.5.2 and later Replacement available in 16.3+
    geo_repositories_synced Yes geo_project_repositories_synced
    geo_repositories_failed Yes geo_project_repositories_failed
    geo_repositories_checksummed Yes geo_project_repositories_checksummed
    geo_repositories_checksum_failed Yes geo_project_repositories_checksum_failed
    geo_repositories_verified Yes geo_project_repositories_verified
    geo_repositories_verification_failed Yes geo_project_repositories_verification_failed
    geo_repositories_checksum_mismatch No None available
    geo_repositories_retrying_verification No None available
    • Impacted versions:
      • 16.3.0 to 16.5.1
    • Versions containing fix:
      • 16.5.2 and later

    For more information, see issue 429617.

  • Object storage verification was added in GitLab 16.4. Due to an issue some Geo installations are reporting high memory usage which can lead to the GitLab application on the primary becoming unresponsive.

    Your installation may be impacted if you have configured it to use object storage and have enabled GitLab-managed object storage replication

    Until this is fixed, the workaround is to disable object storage verification. Run the following command on one of the Rails nodes on the primary site: 

    sudo gitlab-rails runner 'Feature.disable(:geo_object_storage_verification)'

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.4 16.4.0 - 16.4.2 16.4.3
    16.5 16.5.0 - 16.5.1 16.5.2

  • An issue with sync states getting stuck in pending state results in replication being stuck indefinitely for impacted items leading to risk of data loss in the event of a failover. This mostly impact repository syncs but can also can also affect container registry syncs. You are advised to upgrade to a fixed version to avoid risk of data loss.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 16.3.0 - 16.3.5 16.3.6
    16.4 16.4.0 - 16.4.1 16.4.2

  • After Group Wiki verification was added in GitLab 16.3, missing Group Wiki repositories are being incorrectly flagged as failing verification. This issue is not a result of an actual replication/verification failure but an invalid internal state for these missing repositories inside Geo and results in errors in the logs and the verification progress reporting a failed state for these Group Wiki repositories.

    See details of the problem and workaround in issue #426571

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 16.5.0 - 16.5.1 16.5.2

  • You might experience verification failures on a subset of projects due to checksum mismatch between the primary site and the secondary site. The details are tracked in this issue. There is no risk of data loss as the data is being correctly replicated to the secondary sites. Users cloning impacted projects from a Geo secondary site will always be redirected to the primary site. There are no known workarounds at this time. We are actively working on a fix.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 16.6.0 - 16.6.5 16.6.6
    16.7 16.7.0 - 16.7.3 16.7.4

16.3.0

  • Update to GitLab 16.3.5 or later. This avoids issue 425971 that causes an excessive use of database disk space for GitLab 16.3.3 and 16.3.4.

  • A unique index was added to ensure that there’s no duplicate NPM packages in the database. If you have duplicate NPM packages, you need to upgrade to 16.1 first, or you are likely to run into the following error: PG::UniqueViolation: ERROR: could not create unique index "idx_packages_on_project_id_name_version_unique_when_npm".

  • For Go applications, crypto/tls: verifying certificate chains containing large RSA keys is slow (CVE-2023-29409) introduced a hard limit of 8192 bits for RSA keys. In the context of Go applications at GitLab, RSA keys can be configured for:

    You should check the size of your RSA keys (openssl rsa -in <your-key-file> -text -noout | grep "Key:") for any of the applications above before upgrading.

  • A BackfillCiPipelineVariablesForPipelineIdBigintConversion background migration is finalized with the EnsureAgainBackfillForCiPipelineVariablesPipelineIdIsFinished post-deploy migration. GitLab 16.2.0 introduced a batched background migration to backfill bigint pipeline_id values on the ci_pipeline_variables table. This migration may take a long time to complete on larger GitLab instances (4 hours to process 50 million rows reported in one case). To avoid a prolonged upgrade downtime, make sure the migration has completed successfully before upgrading to 16.3.

    You can check the size of the ci_pipeline_variables table in the database console: 

    select count(*) from ci_pipeline_variables;

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

Specific information applies to Linux package installations:

  • In GitLab 16.0, we announced an upgraded base Docker image, which has a new version of OpenSSH Server. An unintended consequence of the new version is that it disables accepting SSH RSA SHA-1 signatures by default. This issue should only impact users using very outdated SSH clients.

    To avoid problems with SHA-1 signatures being unavailable, users should update their SSH clients because using SHA-1 signatures is discouraged by the upstream library for security reasons.

    To allow for a transition period where users can’t immediately upgrade their SSH clients, GitLab 16.3 and later has support for a GITLAB_ALLOW_SHA1_RSA environment variable in the Dockerfile. If GITLAB_ALLOW_SHA1_RSA is set to true, this deprecated support is reactivated.

    Because we want to foster security best practices and follow the upstream recommendation, this environment variable will only be available until GitLab 17.0, when we plan to drop support for it.

    For more information, see:

Geo installations

Specific information applies to installations using Geo:

  • Git pulls against a secondary Geo site are being proxied to the primary Geo site even when that secondary site is up to date. You are impacted if you are using Geo to accelerate remote users who make Git pull requests against a secondary Geo site.

    • Impacted versions:
      • 16.3.0 to 16.3.3
    • Versions containing fix:
      • 16.3.4 and later

    For more information, see issue 425224.

  • A number of Prometheus metrics were incorrectly removed in 16.3.0, which can break dashboards and alerting:

    Affected metric Metric restored in 16.5.2 and later Replacement available in 16.3+
    geo_repositories_synced Yes geo_project_repositories_synced
    geo_repositories_failed Yes geo_project_repositories_failed
    geo_repositories_checksummed Yes geo_project_repositories_checksummed
    geo_repositories_checksum_failed Yes geo_project_repositories_checksum_failed
    geo_repositories_verified Yes geo_project_repositories_verified
    geo_repositories_verification_failed Yes geo_project_repositories_verification_failed
    geo_repositories_checksum_mismatch No None available
    geo_repositories_retrying_verification No None available
    • Impacted versions:
      • 16.3.0 to 16.5.1
    • Versions containing fix:
      • 16.5.2 and later

    For more information, see issue 429617.

  • An issue with sync states getting stuck in pending state results in replication being stuck indefinitely for impacted items leading to risk of data loss in the event of a failover. This mostly impact repository syncs but can also can also affect container registry syncs. You are advised to upgrade to a fixed version to avoid risk of data loss.

  • You might experience verification failures on a subset of projects due to checksum mismatch between the primary site and the secondary site. The details are tracked in issue 427493. There is no risk of data loss as the data is being correctly replicated to the secondary sites. Users cloning impacted projects from a Geo secondary site will always be redirected to the primary site. There are no known workarounds, you should upgrade to a version that contains the fix.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 16.6.0 - 16.6.5 16.6.6
    16.7 16.7.0 - 16.7.3 16.7.4

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 16.3.0 - 16.3.5 16.3.6
    16.4 16.4.0 - 16.4.1 16.4.2

  • After Group Wiki verification was added in GitLab 16.3, missing Group Wiki repositories are being incorrectly flagged as failing verification. This issue is not a result of an actual replication/verification failure but an invalid internal state for these missing repositories inside Geo and results in errors in the logs and the verification progress reporting a failed state for these Group Wiki repositories.

    See details of the problem and workaround in issue #426571

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    16.3 All None
    16.4 All None
    16.5 16.5.0 - 16.5.1 16.5.2

16.2.0

  • Legacy LDAP configuration settings may cause NoMethodError: undefined method 'devise' for User:Class errors. This error occurs if you have TLS options (such as ca_file) not specified in the tls_options hash, or use the legacy gitlab_rails['ldap_host'] option. See the configuration workarounds for more details.

  • If your GitLab database was created by or upgraded via versions 15.11.0 - 15.11.2 inclusive, upgrading to GitLab 16.2 fails with: 

    PG::UndefinedColumn: ERROR:  column "id_convert_to_bigint" of relation "ci_build_needs" does not exist
LINE 1: ...db_config_name:main*/ UPDATE "ci_build_needs" SET "id_conver...

    See the details and workaround.

  • You might encounter the following error while upgrading to GitLab 16.2 or later: 

    main: == 20230620134708 ValidateUserTypeConstraint: migrating =======================
main: -- execute("ALTER TABLE users VALIDATE CONSTRAINT check_0dd5948e38;")
rake aborted!
StandardError: An error has occurred, all later migrations canceled:
PG::CheckViolation: ERROR:  check constraint "check_0dd5948e38" of relation "users" is violated by some row

    For more information, see issue 421629.

  • You might encounter the following error after upgrading to GitLab 16.2 or later: 

    PG::NotNullViolation: ERROR:  null value in column "source_partition_id" of relation "ci_sources_pipelines" violates not-null constraint

    Sidekiq and Puma processes must be restarted to resolve this issue.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

Specific information applies to Linux package installations:

  • As of GitLab 16.2, PostgreSQL 13.11 and 14.8 are both shipped with the Linux package. During a package upgrade, the database isn’t upgraded to PostgreSQL 14. If you want to upgrade to PostgreSQL 14, you must do it manually: 

    sudo gitlab-ctl pg-upgrade -V 14

    PostgreSQL 14 isn’t supported on Geo deployments and is planned for future releases.

  • In 16.2, we are upgrading Redis from 6.2.11 to 7.0.12. This upgrade is expected to be fully backwards compatible.

    Redis is not automatically restarted as part of gitlab-ctl reconfigure. Hence, users are manually required to run sudo gitlab-ctl restart redis after the reconfigure run so that the new Redis version gets used. A warning mentioning that the installed Redis version is different than the one running is displayed at the end of reconfigure run until the restart is performed.

    Follow the zero-downtime instructions for upgrading your Redis HA cluster.

Self-compiled installations

Geo installations

Specific information applies to installations using Geo:

  • New job artifacts are not replicated by Geo if job artifacts are configured to be stored in object storage and direct_upload is enabled. This bug is fixed in GitLab versions 16.1.4, 16.2.3, 16.3.0, and later.
    • Impacted versions: GitLab versions 16.1.0 - 16.1.3 and 16.2.0 - 16.2.2.
    • While running an affected version, artifacts which appeared to become synced may actually be missing on the secondary site. Affected artifacts are automatically resynced upon upgrade to 16.1.5, 16.2.5, 16.3.1, 16.4.0, or later. You can manually resync affected job artifacts if needed.

Cloning LFS objects from secondary site downloads from the primary site

A bug in the Geo proxying logic for LFS objects meant that all LFS clone requests against a secondary site are proxied to the primary even if the secondary site is up-to-date. This can result in increased load on the primary site and longer access times for LFS objects for users cloning from the secondary site.

In GitLab 15.1 proxying was enabled by default.

You are not impacted:

  • If your installation is not configured to use LFS objects
  • If you do not use Geo to accelerate remote users
  • If you are using Geo to accelerate remote users but have disabled proxying
Affected minor releases Affected patch releases Fixed in
15.1 - 16.2 All 16.3 and later

Workaround: A possible workaround is to disable proxying. Note that the secondary site fails to serve LFS files that have not been replicated at the time of cloning.

16.1.0

  • A BackfillPreparedAtMergeRequests background migration is finalized with the FinalizeBackFillPreparedAtMergeRequests post-deploy migration. GitLab 15.10.0 introduced a batched background migration to backfill prepared_at values on the merge_requests table. This migration may take multiple days to complete on larger GitLab instances. Make sure the migration has completed successfully before upgrading to 16.1.0.
  • GitLab 16.1.0 includes a batched background migration MarkDuplicateNpmPackagesForDestruction to mark duplicate NPM packages for destruction. Make sure the migration has completed successfully before upgrading to 16.3.0 or later.

  • A BackfillCiPipelineVariablesForBigintConversion background migration is finalized with the EnsureBackfillBigintIdIsCompleted post-deploy migration. GitLab 16.0.0 introduced a batched background migration to backfill bigint id values on the ci_pipeline_variables table. This migration may take a long time to complete on larger GitLab instances (4 hours to process 50 million rows reported in one case). To avoid a prolonged upgrade downtime, make sure the migration has completed successfully before upgrading to 16.1.

    You can check the size of the ci_pipeline_variables table in the database console: 

    select count(*) from ci_pipeline_variables;

Self-compiled installations

  • You must remove any settings related to Puma worker killer from the puma.rb configuration file, because those have been removed. For more information, see the puma.rb.example file.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Geo installations

Tier: Premium, Ultimate Offering: Self-managed

Specific information applies to installations using Geo:

  • Some project imports do not initialize wiki repositories on project creation. See the details and workaround.
  • Because of the migration of project designs to SSF, missing design repositories are being incorrectly flagged as failing verification. This issue is not a result of an actual replication/verification failure but an invalid internal state for these missing repositories inside Geo and results in errors in the logs and the verification progress reporting a failed state for these design repositories. You could be impacted by this issue even if you have not imported projects.
    • Impacted versions: GitLab versions 16.1.0 - 16.1.2
    • Versions containing fix: GitLab 16.1.3 and later.
  • New job artifacts are not replicated by Geo if job artifacts are configured to be stored in object storage and direct_upload is enabled. This bug is fixed in GitLab versions 16.1.4, 16.2.3, 16.3.0, and later.
    • Impacted versions: GitLab versions 16.1.0 - 16.1.3 and 16.2.0 - 16.2.2.
    • While running an affected version, artifacts which appeared to become synced may actually be missing on the secondary site. Affected artifacts are automatically resynced upon upgrade to 16.1.5, 16.2.5, 16.3.1, 16.4.0, or later. You can manually resync affected job artifacts if needed.
    • Cloning LFS objects from secondary site downloads from the primary site even when secondary is fully synced. See the details and workaround.

Wiki repositories not initialized on project creation

Affected minor releases Affected patch releases Fixed in
15.11 All None
16.0 All None
16.1 16.1.0 - 16.1.2 16.1.3 and later

Some project imports do not initialize wiki repositories on project creation. Since the migration of project wikis to SSF, missing wiki repositories are being incorrectly flagged as failing verification. This is not a result of an actual replication/verification failure but an invalid internal state for these missing repositories inside Geo and results in errors in the logs and the verification progress reporting a failed state for these wiki repositories. If you have not imported projects you are not impacted by this issue.

16.0.0

  • Sidekiq crashes if there are non-ASCII characters in the /etc/gitlab/gitlab.rb file. You can fix this by following the workaround in issue 412767.
  • Sidekiq jobs are only routed to default and mailers queues by default, and as a result, every Sidekiq process also listens to those queues to ensure all jobs are processed across all queues. This behavior does not apply if you have configured the routing rules.
  • Docker 20.10.10 or later is required to run the GitLab Docker image. Older versions throw errors on startup.

  • Container registry using Azure storage might be empty with zero tags. You can fix this by following the breaking change instructions.

  • Normally, backups in environments that have PgBouncer must bypass PgBouncer by setting variables that are prefixed with GITLAB_BACKUP_. However, due to an issue, gitlab-backup uses the regular database connection through PgBouncer instead of the direct connection defined in the override, and the database backup fails. The workaround is to use pg_dump directly.

    Affected releases:

    Affected minor releases Affected patch releases Fixed in
    15.11 All None
    16.0 All None
    16.1 All None
    16.2 All None
    16.3 All None
    16.4 All None
    16.5 All None
    16.6 All None
    16.7 16.7.0 - 16.7.6 16.7.7
    16.8 16.8.0 - 16.8.3 16.8.4

Linux package installations

Specific information applies to Linux package installations:

  • The binaries for PostgreSQL 12 have been removed.

    Prior to upgrading, administrators of Linux package installations must ensure the installation is using PostgreSQL 13.

  • Grafana that was bundled with GitLab is deprecated and is no longer supported. It is removed in GitLab 16.3.

  • This upgrades openssh-server to 1:8.9p1-3.

    Using ssh-keyscan -t rsa with older OpenSSH clients to obtain public key information is no longer viable because of the deprecations listed in OpenSSH 8.7 Release Notes.

    Workaround is to make use of a different key type, or upgrade the client OpenSSH to a version >= 8.7.

  • Migrate your Praefect configuration to the new structure to ensure all your praefect['..'] settings continue to work in GitLab 16.0 and later.

  • Migrate your Gitaly configuration to the new structure to ensure all your gitaly['..'] settings continue to work in GitLab 16.0 and later.

Non-expiring access tokens

Access tokens that have no expiration date are valid indefinitely, which is a security risk if the access token is divulged.

When you upgrade to GitLab 16.0 and later, any personal, project, or group access token that does not have an expiration date automatically has an expiration date set at one year from the date of upgrade.

Before this automatic expiry date is applied, you should do the following to minimize disruption:

  1. Identify any access tokens without an expiration date.
  2. Give those tokens an expiration date.

For more information, see the:

Geo installations

Tier: Premium, Ultimate Offering: Self-managed

Specific information applies to installations using Geo:

Gitaly configuration structure change

The Gitaly configuration structure in the Linux package changes in GitLab 16.0 to be consistent with the Gitaly configuration structure used in self-compiled installations.

As a result of this change, a single hash under gitaly['configuration'] holds most Gitaly configuration. Some gitaly['..'] configuration options continue to be used by GitLab 16.0 and later:

  • enable
  • dir
  • bin_path
  • env_directory
  • env
  • open_files_ulimit
  • consul_service_name
  • consul_service_meta

Migrate by moving your existing configuration under the new structure. git_data_dirs is supported until GitLab 17.0. The new structure is supported from GitLab 15.10.

Migrate to the new structure

caution
If you are running Gitaly cluster, migrate Praefect to the new configuration structure first. Once this change is tested, proceed with your Gitaly nodes. If Gitaly is misconfigured as part of the configuration structure change, repository verification will delete metadata required for Gitaly cluster to work. To protect against configuration mistakes, temporarily disable repository verification in Praefect.
  1. If you’re running Gitaly Cluster, ensure repository verification is disabled on all Praefect nodes. Configure verification_interval: 0, and apply with gitlab-ctl reconfigure.
  2. To apply the new structure to your configuration:
    1. Replace the ... with the value from the old key.
    2. When configuring storage to replace git_data_dirs, append /repositories to value of path as documented below. If you don’t complete this step, your Git repositories are inaccessible until the configuration is fixed. This misconfiguration can cause metadata deletion.
    3. Skip any keys you haven’t configured a value for previously.
    4. Recommended. Include a trailing comma for all hash keys so the hash remains valid when keys are re-ordered or additional keys are added.
  3. Apply the change with gitlab-ctl reconfigure.
  4. Test Git repository functionality in GitLab.
  5. Remove the old keys from the configuration once migrated, and then re-run gitlab-ctl reconfigure.
  6. Recommended, if you’re running Gitaly Cluster. Reinstate Praefect repository verification by removing verification_interval: 0.

The new structure is documented below with the old keys described in a comment above the new keys.

caution
Double check your update to storage. You must append /repositories to the value of path.
gitaly['configuration'] = {
  # gitaly['socket_path']
  socket_path: ...,
  # gitaly['runtime_dir']
  runtime_dir: ...,
  # gitaly['listen_addr']
  listen_addr: ...,
  # gitaly['prometheus_listen_addr']
  prometheus_listen_addr: ...,
  # gitaly['tls_listen_addr']
  tls_listen_addr: ...,
  tls: {
    # gitaly['certificate_path']
    certificate_path: ...,
    # gitaly['key_path']
    key_path: ...,
  },
  # gitaly['graceful_restart_timeout']
  graceful_restart_timeout: ...,
  logging: {
    # gitaly['logging_level']
    level: ...,
    # gitaly['logging_format']
    format: ...,
    # gitaly['logging_sentry_dsn']
    sentry_dsn: ...,
    # gitaly['logging_ruby_sentry_dsn']
    ruby_sentry_dsn: ...,
    # gitaly['logging_sentry_environment']
    sentry_environment: ...,
    # gitaly['log_directory']
    dir: ...,
  },
  prometheus: {
    # gitaly['prometheus_grpc_latency_buckets']. The old value was configured as a string
    # such as '[0, 1, 2]'. The new value must be an array like [0, 1, 2].
    grpc_latency_buckets: ...,
  },
  auth: {
    # gitaly['auth_token']
    token: ...,
    # gitaly['auth_transitioning']
    transitioning: ...,
  },
  git: {
    # gitaly['git_catfile_cache_size']
    catfile_cache_size: ...,
    # gitaly['git_bin_path']
    bin_path: ...,
    # gitaly['use_bundled_git']
    use_bundled_binaries: ...,
    # gitaly['gpg_signing_key_path']
    signing_key: ...,
    # gitaly['gitconfig']. This is still an array but the type of the elements have changed.
    config: [
      {
        # Previously the elements contained 'section', and 'subsection' in addition to 'key'. Now
        # these all should be concatenated into just 'key', separated by dots. For example,
        # {section: 'first', subsection: 'middle', key: 'last', value: 'value'}, should become
        # {key: 'first.middle.last', value: 'value'}.
        key: ...,
        value: ...,
      },
    ],
  },
  # Storage could previously be configured through either gitaly['storage'] or 'git_data_dirs'. Migrate
  # the relevant configuration according to the instructions below.
  # For 'git_data_dirs', migrate only the 'path' to the gitaly['configuration'] and leave the rest of it untouched.
  storage: [
    {
      # gitaly['storage'][<index>]['name']
      #
      # git_data_dirs[<name>]. The storage name was configured as a key in the map.
      name: ...,
      # gitaly['storage'][<index>]['path']
      #
      # git_data_dirs[<name>]['path']. Use the value from git_data_dirs[<name>]['path'] and append '/repositories' to it.
      #
      # For example, if the path in 'git_data_dirs' was '/var/opt/gitlab/git-data', use
      # '/var/opt/gitlab/git-data/repositories'. The '/repositories' extension was automatically
      # appended to the path configured in `git_data_dirs`.
      path: ...,
    },
  ],
  hooks: {
    # gitaly['custom_hooks_dir']
    custom_hooks_dir: ...,
  },
  daily_maintenance: {
    # gitaly['daily_maintenance_disabled']
    disabled: ...,
    # gitaly['daily_maintenance_start_hour']
    start_hour: ...,
    # gitaly['daily_maintenance_start_minute']
    start_minute: ...,
    # gitaly['daily_maintenance_duration']
    duration: ...,
    # gitaly['daily_maintenance_storages']
    storages: ...,
  },
  cgroups: {
    # gitaly['cgroups_mountpoint']
    mountpoint: ...,
    # gitaly['cgroups_hierarchy_root']
    hierarchy_root: ...,
    # gitaly['cgroups_memory_bytes']
    memory_bytes: ...,
    # gitaly['cgroups_cpu_shares']
    cpu_shares: ...,
    repositories: {
      # gitaly['cgroups_repositories_count']
      count: ...,
      # gitaly['cgroups_repositories_memory_bytes']
      memory_bytes: ...,
      # gitaly['cgroups_repositories_cpu_shares']
      cpu_shares: ...,
    }
  },
  # gitaly['concurrency']. While the structure is the same, the string keys in the array elements
  # should be replaced by symbols as elsewhere. {'key' => 'value'}, should become {key: 'value'}.
  concurrency: ...,
  # gitaly['rate_limiting']. While the structure is the same, the string keys in the array elements
  # should be replaced by symbols as elsewhere. {'key' => 'value'}, should become {key: 'value'}.
  rate_limiting: ...,
  pack_objects_cache: {
    # gitaly['pack_objects_cache_enabled']
    enabled: ...,
    # gitaly['pack_objects_cache_dir']
    dir: ...,
    # gitaly['pack_objects_cache_max_age']
    max_age: ...,
  }
}

Praefect configuration structure change

The Praefect configuration structure in the Linux package changes in GitLab 16.0 to be consistent with the Praefect configuration structure used in self-compiled installations.

As a result of this change, a single hash under praefect['configuration'] holds most Praefect configuration. Some praefect['..'] configuration options continue to be used by GitLab 16.0 and later:

  • enable
  • dir
  • log_directory
  • env_directory
  • env
  • wrapper_path
  • auto_migrate
  • consul_service_name

Migrate by moving your existing configuration under the new structure. The new structure is supported from GitLab 15.9.

Migrate to the new structure

caution
Migrate Praefect to the new configuration structure first. Once this change is tested, proceed with your Gitaly nodes. If Gitaly is misconfigured as part of the configuration structure change, repository verification will delete metadata required for Gitaly cluster to work. To protect against configuration mistakes, temporarily disable repository verification in Praefect.
  1. When applying the new structure to your configuration:
    • Replace the ... with the value from the old key.
    • Disable repository verification using verification_interval: 0, as shown below.
    • Skip any keys you haven’t configured a value for previously.
    • Recommended. Include a trailing comma for all hash keys so the hash remains valid when keys are re-ordered or additional keys are added.
  2. Apply the change with gitlab-ctl reconfigure.
  3. Test Git repository functionality in GitLab.
  4. Remove the old keys from the configuration once migrated, and then re-run gitlab-ctl reconfigure.

The new structure is documented below with the old keys described in a comment above the new keys. 

praefect['configuration'] = {
  # praefect['listen_addr']
  listen_addr: ...,
  # praefect['socket_path']
  socket_path: ...,
  # praefect['prometheus_listen_addr']
  prometheus_listen_addr: ...,
  # praefect['tls_listen_addr']
  tls_listen_addr: ...,
  # praefect['separate_database_metrics']
  prometheus_exclude_database_from_default_metrics: ...,
  auth: {
    # praefect['auth_token']
    token: ...,
    # praefect['auth_transitioning']
    transitioning: ...,
  },
  logging: {
    # praefect['logging_format']
    format: ...,
    # praefect['logging_level']
    level: ...,
  },
  failover: {
    # praefect['failover_enabled']
    enabled: ...,
  },
  background_verification: {
    # praefect['background_verification_delete_invalid_records']
    delete_invalid_records: ...,
    # praefect['background_verification_verification_interval']
    #
    # IMPORTANT:
    # As part of reconfiguring Praefect, disable this feature.
    # Read about this above.
    #
    verification_interval: 0,
  },
  reconciliation: {
    # praefect['reconciliation_scheduling_interval']
    scheduling_interval: ...,
    # praefect['reconciliation_histogram_buckets']. The old value was configured as a string
    # such as '[0, 1, 2]'. The new value must be an array like [0, 1, 2].
    histogram_buckets: ...,
  },
  tls: {
    # praefect['certificate_path']
    certificate_path: ...,
   # praefect['key_path']
    key_path: ...,
  },
  database: {
    # praefect['database_host']
    host: ...,
    # praefect['database_port']
    port: ...,
    # praefect['database_user']
    user: ...,
    # praefect['database_password']
    password: ...,
    # praefect['database_dbname']
    dbname: ...,
    # praefect['database_sslmode']
    sslmode: ...,
    # praefect['database_sslcert']
    sslcert: ...,
    # praefect['database_sslkey']
    sslkey: ...,
    # praefect['database_sslrootcert']
    sslrootcert: ...,
    session_pooled: {
      # praefect['database_direct_host']
      host: ...,
      # praefect['database_direct_port']
      port: ...,
      # praefect['database_direct_user']
      user: ...,
      # praefect['database_direct_password']
      password: ...,
      # praefect['database_direct_dbname']
      dbname: ...,
      # praefect['database_direct_sslmode']
      sslmode: ...,
      # praefect['database_direct_sslcert']
      sslcert: ...,
      # praefect['database_direct_sslkey']
      sslkey: ...,
      # praefect['database_direct_sslrootcert']
      sslrootcert: ...,
    }
  },
  sentry: {
    # praefect['sentry_dsn']
    sentry_dsn: ...,
    # praefect['sentry_environment']
    sentry_environment: ...,
  },
  prometheus: {
    # praefect['prometheus_grpc_latency_buckets']. The old value was configured as a string
    # such as '[0, 1, 2]'. The new value must be an array like [0, 1, 2].
    grpc_latency_buckets: ...,
  },
  # praefect['graceful_stop_timeout']
  graceful_stop_timeout: ...,
  # praefect['virtual_storages']. The old value was a hash map but the new value is an array.
  virtual_storage: [
    {
      # praefect['virtual_storages'][VIRTUAL_STORAGE_NAME]. The name was previously the key in
      # the 'virtual_storages' hash.
      name: ...,
      # praefect['virtual_storages'][VIRTUAL_STORAGE_NAME]['nodes'][NODE_NAME]. The old value was a hash map
      # but the new value is an array.
      node: [
        {
          # praefect['virtual_storages'][VIRTUAL_STORAGE_NAME]['nodes'][NODE_NAME]. Use NODE_NAME key as the
          # storage.
          storage: ...,
          # praefect['virtual_storages'][VIRTUAL_STORAGE_NAME]['nodes'][NODE_NAME]['address'].
          address: ...,
          # praefect['virtual_storages'][VIRTUAL_STORAGE_NAME]['nodes'][NODE_NAME]['token'].
          token: ...,
        },
      ],
    }
  ]
}

Disable the second database connection

In GitLab 16.0, GitLab defaults to using two database connections that point to the same PostgreSQL database.

PostgreSQL might need to be configured with a larger value for max_connections. There is a Rake task for checking if this is necessary.

If you have PgBouncer deployed:

Follow the instructions for your installation type to switch back to a single database connection:

Linux package and Docker

  1. Add this setting to /etc/gitlab/gitlab.rb: 

    gitlab_rails['databases']['ci']['enable'] = false

  2. Run gitlab-ctl reconfigure.

In a multi-node environment, this setting should be updated on all Rails and Sidekiq nodes.

Helm chart (Kubernetes)

Set the ci.enabled key to false: 

global:
  psql:
    ci:
      enabled: false
Self-compiled (source)

Remove the ci: section from config/database.yml.

Long-running user type data change

GitLab 16.0 is a required stop for large GitLab instances with a lot of records in the users table.

The threshold is 30,000 users, which includes:

  • Developers and other users in any state, including active, blocked, and pending approval.
  • Bot accounts for project and group access tokens.

GitLab 16.0 introduced a batched background migration to migrate user_type values from NULL to 0. This migration might take multiple days to complete on larger GitLab instances. Make sure the migration has completed successfully before upgrading to 16.1.0 or later.

GitLab 16.1 introduces the FinalizeUserTypeMigration migration which ensures the 16.0 MigrateHumanUserType background migration is completed, executing the 16.0 change synchronously during the upgrade if it’s not completed.

GitLab 16.2 implements a NOT NULL database constraint which fails if the 16.0 migration is not complete.

If 16.0 has been skipped (or the 16.0 migration is not complete) subsequent Linux package (Omnibus) and Docker upgrades might fail after an hour: 

FATAL: Mixlib::ShellOut::CommandTimeout: rails_migration[gitlab-rails]
[..]
Mixlib::ShellOut::CommandTimeout: Command timed out after 3600s:

There is a fix-forward workaround for this issue.

While the workaround is completing the database changes, GitLab is likely to be in an unusable state, generating 500 errors. The errors are caused by Sidekiq and Puma running application code that is incompatible with the database schema.

At the end of the workaround process, Sidekiq and Puma are restarted to resolve that issue.

Undefined column error upgrading to 16.2 or later

A bug in GitLab 15.11 incorrectly disabled a database change on self-managed instances. For more information, see issue 408835.

If your GitLab instance upgraded first to 15.11.0, 15.11.1, or 15.11.2 the database schema is incorrect and upgrading to GitLab 16.2 or later fails with an error. A database change requires the earlier modification to be in place: 

PG::UndefinedColumn: ERROR:  column "id_convert_to_bigint" of relation "ci_build_needs" does not exist
LINE 1: ...db_config_name:main*/ UPDATE "ci_build_needs" SET "id_conver...

GitLab 15.11.3 shipped a fix for this bug, but it doesn’t correct the problem on instances already running the earlier 15.11 releases.

If you’re not sure if an instance is affected, check for the column on the database console: 

select pg_typeof (id_convert_to_bigint) from public.ci_build_needs limit 1;

If you need the workaround, this query fails: 

ERROR:  column "id_convert_to_bigintd" does not exist
LINE 1: select pg_typeof (id_convert_to_bigintd) from public.ci_buil...

Unaffected instances return: 

 pg_typeof
-----------
 bigint

The workaround for this issue differs if your GitLab instance’s database schema was recently created:

Installation version Workaround
15.9 or earlier 15.9
15.10 15.10
15.11 15.11

Most instances should use the 15.9 procedure. Only very new instances require the the 15.10 or 15.11 procedures. If you’ve migrated GitLab using backup and restore, the database schema comes from the original instance. Select the workaround based on the source instance.

The commands in the following sections are for Linux package installations, and differ for other installation types:

Docker
  • Omit sudo

  • Shell into the GitLab container and run the same commands: 

    docker exec -it <container-id> bash
Self-compiled (source)
Helm chart (Kubernetes)

Workaround: instance created with 15.9 or earlier 

# Restore schema
sudo gitlab-psql -c "DELETE FROM schema_migrations WHERE version IN ('20230130175512', '20230130104819');"
sudo gitlab-rake db:migrate:up VERSION=20230130175512
sudo gitlab-rake db:migrate:up VERSION=20230130104819

# Re-schedule background migrations
sudo gitlab-rake db:migrate:down VERSION=20230130202201
sudo gitlab-rake db:migrate:down VERSION=20230130110855
sudo gitlab-rake db:migrate:up VERSION=20230130202201
sudo gitlab-rake db:migrate:up VERSION=20230130110855

Workaround: instance created with 15.10 

# Restore schema for sent_notifications
sudo gitlab-psql -c "DELETE FROM schema_migrations WHERE version = '20230130175512';"
sudo gitlab-rake db:migrate:up VERSION=20230130175512

# Re-schedule background migration for sent_notifications
sudo gitlab-rake db:migrate:down VERSION=20230130202201
sudo gitlab-rake db:migrate:up VERSION=20230130202201

# Restore schema for ci_build_needs
sudo gitlab-rake db:migrate:down VERSION=20230321163547
sudo gitlab-psql -c "INSERT INTO schema_migrations (version) VALUES ('20230321163547');"

Workaround: instance created with 15.11 

# Restore schema for sent_notifications
sudo gitlab-rake db:migrate:down VERSION=20230411153310
sudo gitlab-psql -c "INSERT INTO schema_migrations (version) VALUES ('20230411153310');"

# Restore schema for ci_build_needs
sudo gitlab-rake db:migrate:down VERSION=20230321163547
sudo gitlab-psql -c "INSERT INTO schema_migrations (version) VALUES ('20230321163547');"